Image Processing Reference

In-Depth Information

2 Camera noise

To apply denoising on raw data, we first need a realistic model for the camera noise. There-

fore, we measure the real camera noise in the raw domain based on a series of exposures and

can be performed with any camera, we use the ARRI Alexa camera, as it delivers uncom-

pressed raw data in 16 bit precision. Since the data is uncompressed, we can expect unaltered

measurement results and additionally the individual camera processing steps are known for

The Alexa camera has been developed for motion picture recordings in digital cinema ap-

plications. It has a CMOS sensor with a resolution of 2880 × 1620. In front of the sensor, the

camera has a filter pack composed of an infrared cut-off filter, an ultraviolet cut-off filter and

an optical low-pass filter to reduce aliasing. The CFA, which is located between the filter pack

and the sensor, is a Bayer mask.

mogenous lighting conditions. The noise variance is calculated as the mean of the difference

between these two frames, the corresponding signal value is calculated as the mean over all

the signal values in these frames. The graph in
Figure 1
(a) shows the variance ploted over the

respective mean signal value. The variance of the sensor noise can be approximated by a linear

ied. We observe, however, one difference in the signal region around value 0.1 × 10
4
. The step

in the variance curve is due to a special characteristic of the Alexa sensor, the dual-gain read-

out technology. The sensor read-out of the Alexa provides two different paths with different

ampliication (dual-gain read-out). The low amplified path provides the data for the signal

range starting from 0.1 × 10
4
. The high amplified path saturates in the high signal values, but

for the low signal values it provides a significantly higher signal-to-noise ratio. The read-out

noise (offset of the variance curve) is reduced, thus the dual-gain technology enhances the low

light performance of the camera. The two read-out paths are combined in the region around

the signal value 0.1 × 10
4
, which explains the step in the variance curve.

FIGURE 1
Variance and distribution of the noise in the raw domain (signal values in 16 bit

precision).

The distribution is very similar to a Gaussian distribution. In
Figure 1
(b) the distribution at

signal level 1265 is shown with the Gaussian approximation. The difference between the ap-

proximation and the measured histogram is small, thus we can well approximate the sensor

noise
n
in the raw domain using a Gaussian distribution with signal-dependent variance.

(1)

Search WWH ::

Custom Search